Qwengram-4B

Frozen Qwen3.5-4B plus an R=1 reader at decoder IDX3/IDX11 (human layers 4/12, 12.5%/37.5% of 32 layers), with a learned global early multiplier and linear token-dependent later arbitration. The external Qwen3.8-Flash-Next PLE supplies the memory. The backbone and PLE remain frozen.

The early multiplier is a learned scalar bounded to [0.75, 1.75]. The later multiplier is 0.5 * sigmoid(w @ RMSNorm(h_late) + b), computed per token.

Canonical endpoint: REAL-15M + linear750 (15,000,064 reader tokens; 749,568 calibration tokens). The matched 10M/15M study independently calibrated both readers. Calibrated 15M improves full-validation, all five domains and their mean over calibrated 10M with paired 95% NLL intervals below zero, without a statistically clear benchmark regression. Unresolved accuracy differences do not establish equivalence.

Frozen full-validation perplexity falls 2.784%, from 11.298811 to 10.984210. Raw 15M achieves a larger 2.970% reduction; calibration trades some aggregate LM gain for lower LAMBADA NLL. Matched 500K controls resolve REAL > PERMUTED > RANDOM > DISABLED. See the decision report for all four endpoints, paired intervals, controls and cross-scale provenance.

GGUF files

File	Backbone precision
QwenGram-4B-BF16.gguf	BF16
QwenGram-4B-Q8_0.gguf	Q8_0
QwenGram-4B-Q6_K.gguf	Q6_K
QwenGram-4B-Q4_K_M.gguf	Q4_K_M

All files contain the backbone plus 11 FP32 reader/arbiter tensors. These extension tensors stay FP32 in every precision. reader.safetensors and arbiter.pt contain the same canonical checkpoint separately. The PLE is a required external file. SHA256.json records hashes and sizes.

Required PLE sidecar

Download Ivan Fioravanti's Q4_1 PLE GGUF. Credit for this PLE conversion belongs to Ivan. Its SHA256 is 66db3ab390f4dd5063ecc89cc180f4713898577682347001bf64ab8e328527a1. The approximately 32 GB file is mapped on the host; selected rows are dequantized per token. It does not require a 32 GB GPU allocation.

Build and run

Use the Qwengram llama.cpp fork at commit 3616a858f2326e87ad8b48e1341a4e341b3dad73:

git clone https://github.com/Ninnix/llama.cpp-qwengram.git
cd llama.cpp-qwengram
git checkout 3616a858f2326e87ad8b48e1341a4e341b3dad73
cmake -S . -B build-qwengram-cpu -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_EXAMPLES=ON
cmake --build build-qwengram-cpu -j --target llama-completion
export QWENGRAM_PLE=/path/to/Qwen3.8-Flash-Next-PLE-Q4_1.gguf
build-qwengram-cpu/bin/llama-completion -m /path/to/QwenGram-4B-Q8_0.gguf -p 'The capital of France is' -n 16 -no-cnv -ngl 0

For Vulkan, build with -DGGML_VULKAN=ON. On the tested AMD BC-250, BF16, Q8_0, Q4_K_M with full Vulkan offload (-ngl 99) matched the corresponding CPU eight-token greedy continuation for The capital of France is. These short checks do not establish broad GPU parity. Q6_K differed at full offload but matched with -ngl 33, which keeps the first decoder layer on CPU. Use -ngl 33 for Q6_K on this tested device, or CPU (-ngl 0). The fallback is also a short generation check.

The fork supports the original 0.8B/2B readers at IDX2/IDX8 and the 2560-wide 4B reader at IDX3/IDX11. Hashing, row ordering and reader math are preserved. Stock upstream llama.cpp does not execute this custom reader. MTP and embedding-only inputs are unsupported. Vision has not been validated.

Frozen evaluation

Canonical REAL-15M + linear750, using the original FP8 PLE and frozen Kaggle suite.

Metric	Frozen stock	Canonical Qwengram-4B
Full-validation NLL	2.424697	2.396459
Full-validation perplexity	11.298811	10.984210
General NLL	2.530256	2.498815
Code NLL	1.221217	1.204523
Math NLL	1.212598	1.196319
Scientific NLL	1.894821	1.891064
Multilingual NLL	2.975046	2.950582
Five-domain mean NLL	1.966787	1.948261
LAMBADA-1000 NLL	1.350000	1.336395
LAMBADA-1000 accuracy	65.9%	66.6%
HellaSwag-1000 accuracy	54.8%	55.1%

All five domains improve over stock with paired NLL intervals below zero. LAMBADA NLL improves significantly; HellaSwag and LAMBADA accuracy gains are point estimates with intervals overlapping zero. Training used a single T4 with decoder activation checkpointing, with a 10.029 GiB measured peak. The study and evaluation artifacts preserve the protocol, per-item scores, paired bootstrap results and checkpoint identities. The historical cross-scale table retains the 2B canonical choice frozen when that study began.

GGUF runtime retention

The matched CPU test scores 8,128 tokens from the first 64 consecutive 256-token WikiText-2 raw test chunks, scoring the last 127 tokens per chunk. All runs use eight threads and context/batch/microbatch 256, with no warmup. Reader gain is NLL(stock) - NLL(Qwengram); retention divides each quantized gain by the BF16 gain. Paired 95% intervals use 10,000 resamples of 16 consecutive four-chunk blocks, seed 1234. The external PLE is Ivan Fioravanti's Q4_1 sidecar.

Precision	Stock NLL	Qwengram NLL	Reader gain [95% CI]	Gain retention [95% CI]	Perplexity reduction vs stock
BF16	2.225750	2.198449	0.027301 [0.022650, 0.031917]	100%	2.69%
Q8_0	2.226990	2.199469	0.027521 [0.022892, 0.032133]	100.8% [97.1%, 104.7%]	2.71%
Q6_K	2.230865	2.200943	0.029922 [0.023478, 0.036439]	109.6% [99.4%, 118.6%]	2.95%
Q4_K_M	2.262792	2.231828	0.030964 [0.025884, 0.036537]	113.4% [100.6%, 127.6%]	3.05%

BF16 has the lowest absolute Qwengram NLL in this test. Gain retention measures the added PLE benefit within each precision.

All 441 backbone tensors and nine tokenizer fields match the corresponding stock controls; all 11 extension tensors remain bit-exact FP32. See tensor verification, sidecar validation, generation checks and the matched report. The Q4_1 runtime test and the FP8 frozen evaluation above are separate benchmarks.

The target model revision is 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. Full provenance is in qwengram-4b.json, evaluation/, and runtime/. Study-relative paths in the checkpoint definition resolve in the linked study repository. This is an experimental text-generation release.